A distribution-level statistical framework for reliable pipeline leak detection using multi-domain signal analysis

Early and reliable detection of pipeline leaks is essential for ensuring operational safety and minimizing environmental and economic losses. In practice, however, pipeline monitoring signals are inherently non-stationary and strongly influenced by operating conditions, making robust leak detection challenging, particularly when labeled fault data are scarce or unavailable. This paper presents a statistically grounded, signal-processing-based framework for pipeline leak detection that operates without reliance on machine learning or deep learning models. Pipeline signals are represented using a compact multi-domain feature set integrating time-domain statistics, frequency-domain spectral descriptors, and time-frequency features derived from wavelet packet decomposition. Instead of monitoring pointwise feature deviations, pipeline condition is assessed through distribution-level comparison between feature sets extracted from sliding monitoring windows and a reference distribution constructed under normal operation. Multiple complementary two-sample statistics energy distance, maximum mean discrepancy, and Hotelling's [Formula: see text] statistic are employed to capture distinct aspects of distributional divergence and fused into a unified health indicator (HI). Leak detection is achieved by comparing this indicator against statistically derived thresholds obtained exclusively from normal-condition data, enabling automated detection with controlled false-alarm behavior. The proposed framework is validated using experimental acoustic emission data collected from pipelines conveying gas and water at pressure levels of 13 bar and 18 bar. Results demonstrate stable normal-state behavior with false-alarm rates below 1%, rapid leak detection within 1-2 samples after leak onset, and detection accuracy exceeding 99% across all operating scenarios. Permutation-based statistical significance analysis confirms strong detection confidence with p-values in the range of [Formula: see text] to [Formula: see text], while robustness evaluation across more than 60 parameter configurations demonstrates consistent performance without operating-condition-specific tuning. The proposed approach provides an interpretable, data-efficient, and statistically rigorous solution for practical pipeline leak monitoring.